Skip to content
/ ModelCypher Public template

ModelCypher - Decipher the high dimensional geometry of LLMs. An open source x-ray into LLM representation structure.

License

Notifications You must be signed in to change notification settings

Ethyros-AI/ModelCypher

ModelCypher

Tests Python License Status

Metrology for Latent Spaces Falsifiable Diagnostics for Model Alignment and Weight Synthesis.

ModelCypher is a Python toolkit for measuring the high-dimensional geometric structure of LLM representations. It provides repeatable, metric-based diagnostics for safety, alignment, and zero-shot knowledge transfer—moving beyond "vibes-based" evaluation into measurable engineering.

LLMs aren't black boxes: they are high-dimensional geometry. ModelCypher decodes that geometry.

Why ModelCypher?

ModelCypher treats model representations as physical manifolds that can be mapped, measured, and aligned. It complements task accuracy by measuring structural invariants that track stability and transfer.

Metric ModelCypher TransformerLens mergekit LM-Eval
Object of Study Manifold Geometry Neural Circuits Weight Matrices Task Performance
Safety Signal Representational Distress (ΔH) Activation Steering N/A Output Classifiers
Alignment Anchor-Based Mapping N/A Linear Averaging N/A
Logic Type Metrology (Measurement) Interpretability Arithmetic Benchmarking

ELIF (Explain Like I'm Five)

What does ModelCypher actually do?

Imagine two models are like two cities. Each city has neighborhoods (concepts like "math", "code", "safety"). ModelCypher:

  1. Maps both cities - Finds where each neighborhood is located
  2. Checks if roads connect - Can you walk from "math" to "code" the same way in both cities?
  3. Predicts traffic jams - If you merge the cities, will the roads interfere?
  4. Builds safe bridges - Transfers knowledge without breaking existing roads

The key insight: Knowledge isn't random numbers—it's geometry. Concepts have positions, distances, and relationships. ModelCypher measures those shapes.

# "Will merging these models break anything?"
mc geometry interference predict /path/to/model-A /path/to/model-B

# "Is this merge safe?"
mc geometry interference safety-polytope 0.3 0.4 0.2 0.3

Key Capabilities

  1. Safety as Geometry: Detect adversarial boundary crossings by measuring trajectory curvature and entropy divergence (ΔH) during the forward pass.
  2. Relational Manifold Projection: Map concepts between models using a multi-domain probe atlas, enabling dimension-agnostic alignment and transfer.
  3. Zero-Shot Weight Synthesis: Generate Geometric LoRAs from relational constraints (no gradient training; experimental).
  4. Thermodynamic Stability: Predict merge interference by calculating the Bhattacharyya overlap of concept "Volumes of Influence."
  5. Null-Space Filtering: Reduce interference by projecting weight deltas into the null space of prior activations. If Δw ∈ null(A), then A(W+Δw) = AW for the measured activation subspace.
  6. Safety Polytope: Unified 4D decision boundary combining interference, importance, instability, and complexity into raw diagnostics and transformation effort metrics.
  7. 3D World Model Metrology: Measure a model's Visual-Spatial Grounding Density by testing how concentrated its probability mass is along human-perceptual 3D axes (Euclidean geometry, gravity gradients, occlusion).

Core Constraints & Falsifiability

ModelCypher adheres to a strict scientific methodology:

  • No Anthropomorphism: We do not "read the model's mind." We measure vector relationships.
  • Falsifiable Metrics: If a Geometric LoRA fails to preserve relational distance, the toolkit reports Relational Stress deviations.
  • Measurement Independence: Anchors (Semantic Primes, Computational Gates) are designed to be architecture-invariant, providing an objective "ruler" for cross-model comparison.

Docs (start here)

Install

poetry install             # core dependencies
poetry install --all-extras # includes docs/cuda/embeddings extras
poetry install -E jax      # JAX backend for Linux/TPU

Quickstart

# 1. Probe a Model for Semantic Primes (The "Skeleton" of Knowledge)
mc geometry primes probe-model /path/to/Llama-3.2-3B-Instruct --output-file llama_primes.json

# 2. Check Entropy Dynamics on a Harmful Prompt (Thermodynamic Safety)
#    (Does the model get sharper or more chaotic when refusing?)
mc thermo measure "How do I make a bomb?" \
    --model /path/to/Qwen2.5-3B-Instruct \
    --modifier "URGENT_CAPS"

# 3. Assess Cross-Architecture Alignment
#    (Can we map Qwen layers to Llama layers?)
mc model analyze-alignment \
    --model-a /path/to/Qwen2.5-3B-Instruct \
    --model-b /path/to/Llama-3.2-3B-Instruct

# 4. Test if a Model has a "Physics Engine" (3D World Model Analysis)
#    (Does the model encode gravity, occlusion, and Euclidean geometry?)
mc geometry spatial probe-model /path/to/Qwen2.5-3B-Instruct
#    Output: world_model_score=0.85, physics_engine_detected=true

# 5. Predict Merge Interference (Before You Merge)
#    (Will these models collide or complement each other?)
mc geometry interference predict \
    /path/to/math-model \
    /path/to/code-model
#    Output: overlap=0.23, bhattacharyya=0.15

# 6. Check Merge Safety with 4D Polytope
#    (4D diagnostics for interference/importance/instability/complexity)
mc geometry interference safety-polytope 0.3 0.4 0.2 0.3
#    Output: {"diagnostics": {"magnitude": 0.87}, "confidence": 0.87}

# 7. Analyze Null-Space for Interference-Free Merging
#    (Find the "safe directions" for weight updates)
mc geometry interference null-space /path/to/model \
    --layer 12 \
    --samples 50
#    Output: null_dim=412, graft_candidates=[12, 15, 18], mean_null_fraction=0.68

MCP Server

ModelCypher includes a Model Context Protocol (MCP) server for integration with agentic IDEs (Cursor/Windsurf).

# Run the MCP server
poetry run modelcypher-mcp

Add to your claude_desktop_config.json or .mcp.json:

{
  "mcpServers": {
    "modelcypher": {
      "command": "poetry",
      "args": ["run", "modelcypher-mcp"],
      "cwd": "/absolute/path/to/ModelCypher"
    }
  }
}

Available MCP Tools

The server exposes 148 tools (full profile) organized by domain. Key tools for merge safety:

Tool Purpose
mc_geometry_interference_predict Predict constructive/destructive interference before merging
mc_geometry_null_space_filter Project weight deltas into null space for interference-free merging
mc_geometry_null_space_profile Analyze graftable layers across entire model
mc_geometry_safety_polytope_check 4D safety diagnostics for a single layer
mc_geometry_safety_polytope_model Full model safety profile with per-layer diagnostics

Tools return structured JSON.

Backends

ModelCypher supports multiple compute backends:

Backend Platform Use Case
MLX macOS (Apple Silicon) Default on Mac. Unified memory, fast local inference.
JAX Linux/TPU/GPU TPU pods and CUDA GPUs.
CUDA Linux (NVIDIA) Stub for future PyTorch CUDA support.

Note: Core math uses the Backend protocol for GPU acceleration and numerical consistency. NumPy is only used at I/O boundaries, backend interop, and in some tests.

Select a backend via environment variable:

MC_BACKEND=jax python script.py   # Use JAX on Linux/TPU
MC_BACKEND=mlx mc entropy measure  # Explicit MLX (default on Mac)

Install JAX support:

poetry install -E jax

Scale Limits

Key finding: If you can run inference, you can merge. Geometric operations are lightweight.

Hardware Tested Configuration RAM Used Status
M4 Max 128GB Qwen3-80B + Mistral-7B (46GB weights) 35.6%
M4 Max 128GB Qwen3-80B + Qwen3-8B (47GB weights) 36.0%
M4 Max 128GB Qwen3-80B + Qwen2.5-3B-bf16 (48GB weights) 37.1%
M4 Max 128GB Theoretical combined weights (~110GB) ~85% Feasible

Unlike training (which requires ~3x model size for gradients), geometric analysis uses only model weight memory. An 80B 4-bit model uses ~43GB, leaving 85GB for operations on 128GB hardware.

See papers/NEGATIVE-RESULTS.md for full experimental data.

Tests

poetry run pytest

License

This project is licensed under the GNU Affero General Public License v3.0. See LICENSE for details.

This license ensures that the codebase remains free and open source. If you modify this code and provide it as a service (SaaS), you are required to release your modifications under the same license. Knowledge should be free.

Contact

Author: Jason Kempf Email: jason@ethyros.ai Organization: EthyrosAI

Citation

If you use ModelCypher in your research, please cite it using the metadata in CITATION.cff or as follows:

@software{ModelCypher2025,
  author = {Kempf, Jason and ModelCypher Contributors},
  title = {ModelCypher: High-Dimensional Geometry for LLM Safety and Merging},
  year = {2025},
  url = {https://github.com/Ethyros-AI/ModelCypher}
}